Secure training of deep neural networks by federated learning: Privacy-preserving practice in medical imaging

Federated learning stands out as a vital solution for ensuring the secure training of deep neural networks in medical applications, effectively navigating challenges associated with the sensitive nature of patient data. With a primary focus on privacy, this methodology facilitates model training on decentralized devices, avoiding the centralization of confidential information. By transmitting only model updates, rather than raw data, federated learning not only enhances security measures but also ensures compliance with regulatory standards. This thesis conducts an extensive study on federated learning, employing experiments to evaluate its adaptability in dynamic and heterogeneous medical datasets—utilizing the BloodMNIST, OrganAMNIST, and PathMNIST datasets. Adopting a non-heterogeneous data distribution approach and employing the FedAvg algorithm at the server side for local model update aggregation, we trained three traditional models (one for each dataset) and 18 federated models, varying in data heterogeneity ratios and user sizes (small, medium, and large-scale collaborations). The findings demonstrate that federated models surpass traditional deep learning models, highlighting the effectiveness of federated learning in medical scenarios.